Git ও GitHub-এ AI প্রজেক্ট

Git-এর জন্ম একটি crisis-এ। ২০০২ থেকে Linux kernel BitKeeper (proprietary VCS) ব্যবহার করত — Larry McVoy বিনামূল্যে দিয়েছিলেন kernel community-কে। ২০০৫-এ disagreement → BitKeeper free version withdraw। Linus-এর কাছে দু'টি option: existing OSS VCS-এ স্বীকার, বা নতুন বানানো। তিনি ২য়টি বেছে নিলেন।

চয়ন-এর কারণ — existing tool কেন insufficient:

• CVS/SVN: centralized — single server failure = sad day। Linus-এর "distributed" need।
• Performance: Linux kernel ১M+ line, ১০K+ contributor। CVS minute-level slow।
• Branching: SVN-এ branch heavy operation। Linus-এর "branch should be cheap, free, plentiful" requirement।
• Integrity: SHA1 content addressing — tampering detection।
• Speed: "cheaper than thinking" — local operation network-roundtrip ছাড়া।

Linus-এর design principles:

1. Distributed by default: প্রত্যেকের local-এ full repo।
2. Content-addressable storage: file = SHA1 hash। Same content = same hash।
3. Snapshot, not delta: commit = full snapshot pointer; storage-এ deduplication।
4. Cheap branching: branch = pointer to commit। Free creation।
5. Strong integrity: SHA chain — tampering detect-able।
6. Speed > UX: "complexity is okay if performance demands"।

প্রথম ১০ দিন (April ২০০৫):

• April 3 — design start।
• April 7 — first self-hosted commit ("the actual git project")।
• April 18 — usable enough for kernel।
• June — kernel migration complete।

Initial criticism:

• "Confusing UI" — porcelain commands counter-intuitive।
• "Steep learning curve" — staging area, rebase, detached HEAD।
• "Mercurial easier" — Hg-ও same time-এ design, friendlier UX।

কেন Git tested time:

• GitHub (২০০৮): hosting + UI + community → critical mass।
• Performance: বড় repo-এও fast। Hg performance কম।
• Open Source: permissive license, community-driven evolution।
• Network effect: developer-রা মুভ — tooling মুভ — IDE মুভ। Switching cost বিশাল।
• Mercurial decline: Bitbucket Hg drop (২০২০) — final nail।

২০২৬-এ Git-এর state:

• ৯০%+ open-source project Git।
• SHA1 → SHA256 transition (slow)।
• Partial clone, sparse checkout — mega-monorepo support।
• Improved UX (git switch, restore — porcelain modernization)।
• AI integration (CoPilot, gh CLI AI assist)।

Critical voices:

• "Plumbing vs porcelain" — internal model leaking through commands।
• Submodule ভয়ঙ্কর UX।
• Merge conflict resolution often painful।
• Pijul, jujutsu — newer attempts at better DAG model।

বৃহত্তর পাঠ:

• Crisis-driven innovation: Linus 10-day produce — necessity strong motivator।
• Right primitive selection: content-addressable + snapshot model — durable foundation।
• Network effect bites both ways: Git-এর alternative replace করা প্রায় অসম্ভব এখন।
• UX matters less than power initially: Git-এর confusing UX-ও tolerated কারণ underlying model strong।

মূল উপলব্ধি: Git-এর সাফল্য = right design + right time + right ecosystem। Linus-এর engineering judgment legendary। ২০ বছর পরেও — ML community নতুন ছাত্র Git শিখছে। Tool নয় — paradigm। Distributed, snapshot-based, content-addressable — যা VCS-এর "operating system"। সম্ভবত বদলাবে; কিন্তু আগের ১০ বছর চলবে।

ML experiment = একটি hyperparameter set + একটি training run + একটি result। ১০ experiment ঠিক, ১০০ — মনে রাখা অসম্ভব। Tracking নাহলে — "তিন সপ্তাহ আগে যে ৯৩% accuracy পেয়েছিলাম, সেই config কী ছিল?" সমস্যা।

চারটি tool, চারটি concern:

(১) Git — code & config:

• What it does: code, config, README — text version control।
• Strength: universal, free, distributed, ৩০-year proven।
• Weakness: large file weak, dataset/model weak, experiment metadata নেই।
• ML use: source code, requirements.txt, Dockerfile, hyperparameter YAML।

(২) DVC — data & model:

• What it does: large file via cloud storage, version-tracked।
• Strength: Git-এর extension, S3/GDrive-agnostic, pipeline definition।
• Weakness: experiment comparison UI সীমিত, web dashboard নেই।
• ML use: dataset versioning, trained model checkpoint, pipeline reproducibility।

(৩) MLflow — experiment lifecycle:

• What it does: log metric/param/artifact, model registry, deployment।
• Strength: open-source, self-hostable, framework-agnostic, ৪ component (tracking, projects, models, registry)।
• Weakness: UI dated, scaling এ heavy server।
• ML use: training run track, A/B compare, prod registry।

(৪) Weights & Biases (W&B):

• What it does: experiment tracking, hyperparameter sweep, dataset/model artifact, report sharing।
• Strength: best-in-class UI, real-time dashboard, team collaboration, integration ১০০+ framework।
• Weakness: SaaS — enterprise-এ pricing climbs; data leaving organization।
• ML use: research labs, model debugging, paper figure generation।

সাধারণ workflow:

# Code & config — Git
git checkout -b experiment/lr-tuning

# Data — DVC (already setup)
dvc pull                   # latest data

# Run experiment — log to W&B (or MLflow)
import wandb
wandb.init(project="iris-classification",
           config={"lr": 0.01, "model": "logreg"})
# ... training ...
wandb.log({"accuracy": 0.97})
wandb.save("model.pkl")    # artifact upload

# Best result-এ Git commit + push
git add .
git commit -m "Best config: lr=0.01, accuracy=0.97"
git push

# Model registry-এ promote
mlflow models register -m wandb-uri/best-run/model -n iris-prod

সিদ্ধান্ত matrix:

• Solo learner: Git only।
• Solo ML practitioner: Git + W&B (free tier)।
• Small team: Git + DVC + MLflow।
• Research lab: Git + DVC + W&B।
• Enterprise: Git + DVC + MLflow + own model registry।

আধুনিক alternatives:

• Comet ML: W&B-র closest competitor।
• Neptune.ai: metadata management focus।
• Aim: open-source W&B alternative।
• ClearML: open-source full lifecycle।
• SageMaker / Vertex AI: cloud-managed end-to-end।
• Hugging Face Hub: model + dataset versioning, free tier generous।

Common antipattern:

• Excel-এ experiment log — copy-paste error।
• Google Sheet shared — no automation, manual update।
• Pickle local-এ — laptop crash = সব lost।
• "Notebook-এ output cell-ই my log" — Restart Run All-এ মুছে যায়।

Discipline tips:

• প্রতিটি run-এ unique ID (W&B/MLflow auto)।
• Hyperparameter config file-এ — code-এ hardcode না।
• Random seed log।
• Git commit hash log — কোন code এই run-এ।
• Dataset version log — কোন data।
• Hardware (GPU type, RAM) log।

মূল উপলব্ধি: ML experiment tracking = scientific notebook-এর digital version। ছোট কাজে — Git + simple log যথেষ্ট। Mid-size কাজে — W&B অসামান্য productivity। Production-এ — MLflow + DVC + registry। Tool বাছার চেয়ে — discipline বেশি গুরুত্বপূর্ণ। লগ লেখা = science, না লেখা = guesswork।

Multi-developer ML team-এ — branching strategy ও merge philosophy পুরো workflow-কে define করে। দু'টি দল আছে: rebase camp ও merge camp। Religious war-ই বটে।

Merge — preserves history:

git checkout main
git merge feature/knn   # creates merge commit

• Branch-এর full history main-এ।
• Merge commit visible — audit trail।
• Non-destructive — original commits unchanged।
• "True history" preserve।
• Cons: history non-linear, "spaghetti" graph।

Rebase — linear history:

git checkout feature/knn
git rebase main         # replay branch on top of main
git checkout main
git merge feature/knn   # fast-forward

• Linear, clean history।
• Each commit-এ branch-এর "ideal" history।
• git log readable।
• Cons: rewrites commit hash; shared branch-এ never rebase।

Squash merge — single commit:

# GitHub PR "Squash and merge" button

• Branch-এর সব commit → main-এ ১টি commit।
• Main super clean; ১ feature = ১ commit।
• Cons: branch-এর fine-grained history lost।

Standard approach:

• Open source / public — merge (preserves contributor history)।
• Internal team — squash (clean main, branch unimportant)।
• Solo / personal — rebase (clean linear)।

Branching strategy — Git Flow:

• main: production releases।
• develop: ongoing development।
• feature/*: new feature off develop।
• release/*: release candidate prep।
• hotfix/*: production emergency।

Vincent Driessen ২০১০-এ propose; ২০১৫ পর্যন্ত popular।

• Pros: clear roles, release-cadence support।
• Cons: complex, slow, ৬টা branch type heavy।

Trunk-based development (২০২০+):

• একটাই long-lived branch — main (trunk)।
• Feature branch ≤১-২ দিন।
• Daily merge to main।
• Feature flag-এ incomplete feature hidden।
• CI/CD → main commit-এ auto deploy।

Pros:

• Faster integration — merge conflict ছোট।
• Continuous delivery natural।
• Branch baggage কম।
• Google, Facebook, Spotify — সব trunk-based।

Cons:

• CI/CD discipline mandatory।
• Test coverage strong লাগে — main always green।
• Feature flag complexity।
• Junior developer-এ direct main commit risky।

GitHub Flow — middle ground (most popular):

• main = production।
• প্রতিটি feature = branch off main → PR → review → merge।
• Branch life ১-৭ days।
• সরল, repeatable।

ML-specific challenge:

• Notebook merge conflict — text not friendly। nbdime tool।
• Long-running experiments — branch দীর্ঘ। Push intermediate result।
• Hyperparameter sweep — branch explosion। DVC pipeline + experiment tracking।
• Model file LFS conflict — usually overwrite latest।

Conflict avoidance practical tips:

• Frequent git pull --rebase — small conflict early।
• Small PR — conflict probability কম।
• Modular code — different file → different developer।
• Pre-merge: git pull origin main && rebase।
• Communication — "এই file আমি touch করছি"।

সমাধান-এর tools:

• git mergetool + meld/kdiff3 — visual merge।
• VS Code merge editor — three-way merge UI।
• nbdime — notebook semantic merge।
• GitHub Codespaces — same branch দেখা।

সমাধানের সিদ্ধান্ত:

• Solo / learning: rebase, simple।
• Open source: merge, preserve contribution history।
• Small team: GitHub Flow + squash merge।
• Mature team: Trunk-based + feature flag।
• Regulated industry: Git Flow with strict review।

মূল উপলব্ধি: Branching strategy — team-এর culture-এর mirror। Strict process slow + safe; fast process risky + agile। Modern trend trunk-based + feature flag — কারণ CI/CD-র সাথে natural। কিন্তু premature adoption — disaster। প্রথমে discipline, পরে strategy।

২০২৬-এ — GitHub profile = ML candidate-এর primary CV-equivalent। LinkedIn experience claim, GitHub demonstrate। Bangladesh থেকে international remote job বা grad school admission — দু'টোতেই key signal।

Recruiter ৩০ সেকেন্ডে যা দেখে:

1. Profile README — kind of human এই person?
2. Pinned 6 repo — best work?
3. Commit graph — consistency, recent activity?
4. Top language — claimed expertise match?
5. Star count one repo — community recognition?

প্রতিটি element-এর strategy:

(১) Profile README:

• Short bio: who, what, where।
• Tech stack — concrete tool, not buzzword soup।
• Current focus — "exploring transformer-based summarization"।
• Open to opportunity — remote ML role, etc।
• Contact — email, LinkedIn।
• Avoid: emoji explosion, fake "fancy" badges।

(২) Pinned repos — quality over quantity:

• ৪-৬টি যা showcase। ৪০টা half-baked-এর চেয়ে ৩টা polished।
• Each repo: clear README, demo screenshot/GIF, run instructions, deployed link if web app।
• Mix: ১টি ML project, ১টি tooling/library, ১টি contribution to open source।

(৩) ML project quality bar:

• Real problem: tutorial Iris-এর বাইরে। Bangladesh-relevant data ভাল।
• Reproducible: requirements.txt + clear setup। Recruiter ৫ minute-এ run করতে পারবেন।
• Documented: README — problem, approach, result, learning।
• Visualization: EDA plot, confusion matrix, ROC curve।
• Honest result: "best-only" না — failure mode জানান।
• Bonus: deployed (Streamlit/HF Spaces), tests, CI badge।

(৪) Bangladesh-relevant project ideas:

• Bengali sentiment analysis — newspaper headlines।
• Bangla handwritten digit recognition (Ekush dataset)।
• Dhaka traffic prediction।
• BD Stock Market analysis।
• Pathao/Foodpanda-style ETA prediction।
• BB exchange rate forecasting।
• BBS data — district-wise socioeconomic clustering।
• Crop disease classification (rice, jute)।
• Bangla ASR (speech recognition) fine-tune।
• Local hospital demand prediction।

(৫) README template — ML project:

# Project Title

One-line elevator pitch.

## 🎯 Problem
What problem this solves, why it matters.

## 📊 Dataset
Source, size, license. Reproducibility note.

## 🔬 Approach
- Preprocessing: ...
- Model: ...
- Why this model: ...

## 📈 Results
| Metric | Value |
|--------|-------|
| Accuracy | 0.92 |

[Confusion matrix image]

## 🚀 Run It Yourself
```bash
git clone ...
pip install -r requirements.txt
python main.py
```

## 💡 Learnings
What I'd do differently next time.

## 🛠 Tech Stack
Python, pandas, scikit-learn, ...

(৬) Open source contribution:

• scikit-learn, pandas, hugging face — beginner-friendly issue।
• Bengali NLP project — direct contribution।
• Documentation fix — easy first PR।
• Translation — Bengali README।
• Hugging Face dataset — Bengali corpus upload।
• "good first issue" tag search।

(৭) Commit graph health:

• Daily small commit > weekly big commit।
• Public commit থাকতে hot — settings-এ private contribution-ও show।
• Authentic — fake automated commit detect হয়।

(৮) Anti-pattern — যা avoid:

• "my-first-tutorial" এ ৪০টা repo।
• Forked repo যেখানে কিছু change নেই।
• README-হীন project।
• Notebook output না clean — bloated commit।
• Empty repo — পুরোনো plan।
• Sensitive data leak — secret commit।

(৯) Bangladesh-specific tactics:

• BUET, DU, BRAC University-র CS lab GitHub organization-এ contribute।
• Local hackathon (BdAPPS, ICPC) winning project — properly document।
• Bengali README + English — bilingual signal।
• Local mentor (Facebook AI Bangladesh group) থেকে review।
• Kaggle competition — link from GitHub।
• Hugging Face Hub থেকে cross-link।

(১০) Time investment realistic:

• ৩ মাস — basic profile + ৩ project।
• ৬ মাস — polished portfolio।
• ১ বছর — open source contribution + recognized work।
• প্রতি সপ্তাহে ৫-১০ ঘণ্টা invested।

মূল উপলব্ধি: GitHub portfolio = scientific habit + creative output + technical execution। Quantity-র চেয়ে quality। Bangladesh-relevant problem + global-quality execution = unique competitive position। Junior থেকে junior-er distinguish — portfolio-ই। Resume বানানো আগে — GitHub plant করুন, water দিন, growth আসবেই।